If you joined a growth team after April 2021, you may not realize how strange the current state of attribution is. We are running paid acquisition with a measurement layer that we know to be partially blind, partially fictional, and partially overstating itself, and we have collectively agreed to behave as if this is fine. It is not fine. It is just survivable.
What broke, exactly
ATT did not destroy attribution. It destroyed device-level user identity for a meaningful share of iOS traffic, which is the substrate that mobile measurement partners (MMPs) had been quietly relying on for half a decade. SKAdNetwork did not replace it; it replaced a small fraction of it with delayed, postback-only, sometimes-aggregated reporting that is genuinely useful for prospecting and almost useless for incremental decisions on a campaign smaller than ~$50k/month. Web tracking was hit by a related but distinct shock: ITP on Safari, third-party cookie deprecation on Chrome (delayed but coming), and the GDPR/LGPD consent waterfall arriving in the same window. The compounded effect: the share of conversions that arrive at your warehouse with a clean session-to-source mapping has dropped from ~95% (2018) to somewhere between 55% and 80% (2026), depending on your traffic mix.
What teams quietly did instead
The honest pattern, observed across maybe forty growth conversations in 2024–2025: teams stopped trusting their MMP last-touch report, stopped raising it in board meetings, and substituted a blended view — usually media-mix modeling (MMM) on a quarterly cadence, plus a 'top-of-funnel signal' (CTR, CPM trends, video-completion) as a real-time proxy. They did not communicate this substitution upward. The CFO is still being shown a CAC number that is dressed in the language of last-touch but actually arrived through a sequence of judgment calls. This is not malicious; it is what survival looks like.
What is actually recoverable
The first useful move is to stop arguing about deterministic recovery and accept that probabilistic + warehouse-grounded signal is the new floor. Concretely: server-side event collection on web (so ITP cannot drop your conversion events), CAPI / Conversion API on the platforms that offer it (Meta, TikTok, Google), and warehouse joining via deterministic IDs (email hash, login event, order ID). The goal is not to reach 100% session-to-source clarity again — that is gone — but to reach 90% clarity on the ~70% of conversions that involve a logged-in or post-login moment, and accept the remaining ~30% as probabilistic noise. Once that floor exists, MMM becomes useful (it has a clean dependent variable to model) and incrementality testing becomes feasible (you have a control arm that is actually comparable).
What to stop arguing about
Three debates we recommend retiring inside the team: (1) 'last-touch vs first-touch' — both are arbitrary picks from a multi-touch path; the right move is to model the path or to test for incrementality, not to pick a single touch. (2) 'attribution model X says we should spend more on Y' — without an incrementality control, attribution model output is opinion, not evidence. (3) 'we lost 18% of conversions to ITP' — you did not lose them; they happened, and your warehouse can recover most of them via post-login deterministic joins, if you have invested in that plumbing.
What we believe
Performance measurement after iOS 14.5 is not a tracking problem. It is a posture problem. Teams that run a clean stack — server-side events, CAPI integrations, warehouse as source of truth, MMM on a cadence, periodic incrementality holdouts — can answer attribution questions with the same confidence as in 2018, just on a slower timeline (weeks, not minutes). Teams that did not invest in that plumbing are running on vibes and last-touch, and have been quietly losing the argument with their CFO for three years. The reconstruction is unglamorous, takes 8–12 weeks, and unlocks every other measurement decision the team will need to make for the next half-decade. It is the single highest-leverage rebuild we know of in the post-ATT period.
What we run
We rebuild this stack as standard for clients on engagements over six months. The order of operations: (1) audit existing event collection and source-of-truth gaps; (2) install server-side event collection and CAPI bridges; (3) wire warehouse joins via deterministic IDs; (4) stand up MMM with a clean baseline; (5) start incrementality testing on the top three channels. The first four steps run in parallel through weeks 1–8; the fifth typically begins in week 9 once the warehouse signal is stable. The deliverable is not a dashboard. It is the ability to answer the next attribution question without flinching.